Computer-aided pulmonary image analysis in small animal models

Abstract

Purpose:

To develop an automated pulmonary image analysis framework for infectious lung diseases in small animal models.

Methods:

The authors describe a novel pathological lung and airway segmentation method for small animals. The proposed framework includes identification of abnormal imaging patterns pertaining to infectious lung diseases. First, the authors’ system estimates an expected lung volume by utilizing a regression function between total lung capacity and approximated rib cage volume. A significant difference between the expected lung volume and the initial lung segmentation indicates the presence of severe pathology, and invokes a machine learning based abnormal imaging pattern detection system next. The final stage of the proposed framework is the automatic extraction of airway tree for which new affinity relationships within the fuzzy connectedness image segmentation framework are proposed by combining Hessian and gray-scale morphological reconstruction filters.

Results:

133 CT scans were collected from four different studies encompassing a wide spectrum of pulmonary abnormalities pertaining to two commonly used small animal models (ferret and rabbit). Sensitivity and specificity were greater than 90% for pathological lung segmentation (average dice similarity coefficient > 0.9). While qualitative visual assessments of airway tree extraction were performed by the participating expert radiologists, for quantitative evaluation the authors validated the proposed airway extraction method by using publicly available EXACT’09 data set.

Conclusions:

The authors developed a comprehensive computer-aided pulmonary image analysis framework for preclinical research applications. The proposed framework consists of automatic pathological lung segmentation and accurate airway tree extraction. The framework has high sensitivity and specificity; therefore, it can contribute advances in preclinical research in pulmonary diseases.

1. INTRODUCTION

Tuberculosis and viral infections not only cause a significant proportion of deaths, but also represent a source of catastrophic pandemics. For example, the 1918 “Spanish” influenza pandemic infected one third of humans and was fatal in 2.5% of cases.¹ Re-emergence of related viral strains, such as pandemic influenza A (H1N1) virus (pH1N1), serves as a reminder that preparedness for such diseases is essential. To counter the burden of disease, and prevent future outbreaks, an improved understanding of their pathogenesis is required. For this purpose, small animal models play an important role in enhancing our understanding of infectious lung diseases. The selection of an appropriate small animal model is dependent on whether the model recapitulates the relevant phenotypes and the availability of appropriate measures of disease. Two examples of this are TB and pH1N1, in which pathogenesis is better recapitulated in rabbits (Oryctolagus cuniculus)^2–4 and ferrets (Mustela putorius furo), respectively.⁵

To obtain phenotype information of small animal models, noninvasive imaging techniques are often used. Although computed tomography (CT) is the most widely available and cost-effective modality for visualizing, analyzing, detecting, and quantifying lung diseases, when performed manually, the quantification process is often time-consuming and subject to user-variability, inaccuracy, and bias.⁶ Therefore, automated computer algorithms are required to provide an option for reducing these discrepancies. To evaluate the disease severity, computerized pulmonary image analysis involves two common tasks: pathological lung segmentation and airway extraction, both of which help radiologists in their diagnostic decisions.

Despite an ever increasing need for robust and reliable delineation techniques, the methodology for automatic analysis of the anatomy and physiology of small animals models is still under development.^6–8 For example, current lung delineation algorithms work well for certain lung abnormalities when present in moderate amounts, but fail to perform when dense or complex pathologies exist. Our aim in this study is to address these challenges, and to develop an automated method for pulmonary CT analysis in small animal infectious disease models. In particular, we develop a robust and reliable tool for segmenting pathological lungs, and extracting airway trees in the presence of abnormalities.

1.A. Our contribution

In this study, we present a pulmonary image analysis and quantification framework that includes a novel pathological lungs and airway segmentation (PLAS) system. To the best of our knowledge, this is the first fully automated computational tool for the analysis of pulmonary infection in small animal models. Experimentally, we study various small animal models from diverse sources containing different types of abnormalities. In order to make our method computationally efficient, we propose a pathology recognition system that reserves the use of the refinement steps for dense abnormalities. A unified platform is intended to facilitate the pulmonary image analysis particularly for infectious lung diseases.

1.B. Related work

Although the literature in small animal disease models using radiological images is vast, most of the pulmonary image analysis methods used in those studies were based on either manual or semiautomated interpretation of the images with limited reproducibility. In particular, there are limited number of pulmonary image analysis systems available for the study of small animal models that analyze lung pathologies and airways in the same framework with high accuracy and efficiency. For a similar study to our framework, for instance, semiautomated analysis of MRI was used to analyze the response to therapy of lung cancers using mice models in Ref. 9, and reproducible quantitative methods were reported to better visualize and understand the severity and progression of the tumors with open source code. For another example, a quantitative image analysis method in longitudinal PET scans was used in Ref. 10 to measure tumor metabolic activity with the total lung volumetry from CT. However, it should be noted that these methods are not only time-consuming due to the semiautomated nature for the image analysis framework but also they are very specific to preclinical application and imaging modalities. Hence, it may not be easy to apply those methods to other animal models. Last but not least, these methods are not scalable in terms of spectrum of the lung pathologies, so there is no module to analyze a wide range of pathology pertaining to lungs.

Threshold-based methods^11–15 are often employed for segmentation and quantification of lungs in small animal model. Such methods are limited in applicability to clinical and preclinical environments as they fail to take into account the intensity variations due to existing pathologies and the structural variability of the organs. Instead, region-based methods such as region growing,¹⁶ random-walk,¹⁷ and fast-marching level-set¹⁸ are often used in handling the intensity variations. These strategies are insufficient, however, if extensive pathologies are present in the scan, as intensity alone may not be sufficient to delineate the lungs. More advanced techniques such as anatomical shape models and atlas-based methods^19,20 have been proposed and applied to human scans with moderate success. In small animal models, these methods fail too in accommodating finer details and usually there is a need for postprocessing steps.¹⁹ Moreover, there is no single CAD system which can work for a wide spectrum of pathologies.²¹

Delineation of pathological areas of lung has traditionally been performed manually or semiautomatically. Recently, local descriptors have been used in CT scans of human subjects to segment localized abnormal regions in the lung field. These approaches identify pathologies through extracting localized discriminating information such as shape and texture. A number of local descriptors for detection of lung pathologies have been proposed: 3D adaptive multiple feature method (AMFM),²² texton-based approach,²³ intensity-based features,²⁴ gray level co-occurrence matrix (GLCM),²⁵ wavelet and Gabor transform,²⁶ shape and context-based attributes,^27,28 local binary patterns (LBP),²⁹ and histogram of gradients (HOG).²⁷ However, the most challenging aspect of these approaches is the selection of descriptors appropriate for the task, which is still an active area of research.

Few strategies address the specific challenges for small animal models, which feature simpler structures, significant motion artifacts, and higher contrast as compared with human scans. Therefore, in order to successfully segment airways from small animal CT scans, modifications to the conventional methods are often necessary. Intensity-based 3D region growing (RG), as a most basic and efficient technique, is commonly used for reliable extraction of large airways including trachea and principal bronchi, but is subject to leakage into lung parenchyma when reaching smaller airways. For human data, such leakage is often caused by similar intensity values and blurred or broken walls. So leakage control can be applied by adaptively selecting proper thresholds³⁰ based on the growing criterion. It is often nontrivial to find optimal threshold values due to motion artifacts as well as poor image resolution. Also, conventional leakage control is limited in improving the overall accuracy. To address this challenge, higher-level information and priors regarding the shape and appearance of airways in CT images can be used. In the literature, while some studies use shape information only, others use sole appearance information. The former works better at extracting the small airways and the latter are suitable for the large airways. Herein, we proposed a novel combination of the two approaches that are intended to improve the identification of both large and small airway delineations. With the contemporary a priori information, locations of finer airways in small animal images can be identified without the leakage problem.

2. MATERIALS AND METHODS

2.A. Overview of the methodology

Figure 1 shows an overview of our PLAS system for small animal models. The PLAS framework consists of three steps: (A) the initial fuzzy connectivity (FC)-based segmentation for normal lung parenchyma extraction, (B) the refinement step triggered by an intelligent pathology detection methodology, and (C) the airway lumen segmentation step. The details of each step are presented in Secs. 2.B–2.D.

FIG. 1.

A block diagram of the PLAS system. GGO: ground glass opacity, Cons.: consolidation.

2.B. Normal lung extraction/initial segmentation

The segmentation of normal lung parenchyma was performed using the FC segmentation framework.^31,32 As illustrated in Fig. 2, FC segmentation requires two seed points: s_l, s_r, located within the left and right lungs, respectively. FC describes the “hanging togetherness” between seed points and all other voxels by examining the path strength between them; and segmentation can be obtained by thresholding over the strength of connectedness.^31,32 In our design, seed locations were automatically identified through a preprocessing step. For a given CT image I, we used a thresholding operation $\mathcal{T}$ based on CT attenuation values for normal lung parenchyma [Hounsfield units (HU): −700 through −400, mean ≈ − 550 HU]. Thus, the thresholded image can be represented as $I^{\mathcal{T}}={\left\{I\right\}}_{-1024}^{-550}$. Note that the purpose of the thresholding operation was to determine lung fields (roughly); hence, the thresholded image included mainly the air spaces. Finally, we set the seed locations s_l and s_r after randomly sampling a few seed candidates (i.e., 3 × 3 × 3 voxel seed windows) from air spaces of both lungs, $I^{\mathcal{T}}$, and select the voxels with minimum HU value as seeds. Note also that we carried out seed sampling within the extracted normal lung parenchyma, not on the original image.

FIG. 2.

Flowchart outlining the initial segmentation using FC.

In addition to initial seeds, the FC segmentation algorithm requires intensity-based-properties such as mean m and the intensity variations parameters σ_ψ and σ_ϕ of lung regions to be used in its affinity functions. These values were empirically set to normal lung parenchyma as m ≈ − 550 HU, σ_ψ = σ_ϕ ≈ 150 HU. These parameters were specifically used in affinity function definitions where Gaussians were used to handle both intra- and interobject variation. Due to space constraints, we refer readers for further technical details of the FC method in Refs. 31 and 32 and references therein. Once the seeds and affinity parameters for FC were set, the output was a binary mask of the lung fields containing the external airways, and the left and right lungs. Figure 3 illustrates how the performance of the FC segmentation may deteriorate with an increase in the extent of pathology (arrows). It should also be noted that although the FC segmentation is more robust than region growing, graph-cut, and other region-based segmentation methods,³³ further refinement is often necessary for cases with dense pathologies.

FIG. 3.

CT images with varying pathologies along with overlaid initial FC segmentation. The arrows point to the lung field regions where FC segmentation failed.

2.C. Refinement of lung segmentation through machine learning classification

2.C.1. Pathology presence test

In order to make our design fully automated and computationally efficient, an intelligent switch-scheme, named the pathology presence test, was included to initiate further refinement to the initial segmentation when necessary. The test consisted of two constraints:

• (i)the smoothness of the lung boundaries and
• (ii)the volume difference between segmented and estimated lung volume,

where estimation of lung volume was calculated from the structural relationship between lung fields and rib cage. Since the rib cage tightly enclose the lung field inside the body, it can provide an approximate value for expected lung volume.

2.C.1.a. The smoothness test.

In general, boundaries of a healthy/normal lung make a smooth transition among the axial slices of a CT scan. On the contrary, segmented lung areas tend to change abruptly between slices where the abnormalities are present. In our work, this smooth transition of the boundary was used as one of the evaluation metrics for the goodness of segmentation. Let A(i) denote the lung area on a given axial slice i, then the change in segmented area in the direction orthogonal to the slice plane can be monitored by C(i) = |A(i + 1) − A(i)|, for i = 1, …, N − 1. To quantify how segmented regions change when abnormality exists in the CT images, we selected a set of control images $C_{\text{ref}}^j$, j = 1,  2, …, M, and calculated the maximal variation for each control image: $K\left(i\right)={max}_j\left(\mathrm{var}\left(C{\left(i\right)}_{\text{ref}}^j\right)\right)$, where var indicates variance operator. The control images were pulmonary CT scans of healthy animals. The scans were selected by the experts through qualitative evaluation. Based on the smoothness assumption, any abrupt change in the segmented area, C(i) > K(i), indicated the presence of a potential abnormality triggering the refinement stage.

2.C.1.b. Volume difference test.

Since rib cage tightly encloses the lungs in small animals, it can be used as an anchor to approximately estimate the enclosed volume when no or minor pathology exists. In fact, estimated lung volume was used as an additional constraint for predicting the existence of pathology. In this work, we used a convex-hull fitting approach around the rib cage to roughly estimate the lung volume.

Fitting a convex-hull to the rib cage structures requires the rib structures to be extracted. Since bone has relatively high HU compared to surrounding tissues, thresholding allows identification of such structures. We chose the thresholding parameter of >300 HU for segmenting all bone structures as convention. Thresholding was followed by a connected component analysis to retain the largest component of the bones and remove noise. Further details on separating adjacent bones from rib cage can be found in our relevant publication.³⁴

A 2D convex-hull, {H(i), i = 0, …, N − 1}, was fitted over the detached scapula in a slice-by-slice manner along the axial plane to obtain rough estimation of the lung field volume ${\hat{V}}_{\mathrm{lf}}$ (Fig. 4). This process can be shown mathematically as

$${\displaystyle {\hat{V}}_{\mathrm{lf}}=f\left\{\sum _{i=0}^{N-1}{V}_{\mathrm{hull}}\left(i\right)\right\},}$$ (1)

FIG. 4.

2D convex-hull (shown in enclosed curve) fitting over the rib cage (boundary bones are highlighted) along the axial plane for pathology presence test of the segmentation.

where f{} is a linear regression function mapping the convex-hull volume, V_hull, to the estimated lung field volume ${\hat{V}}_{\mathrm{lf}}$. Note that ${\hat{V}}_{\mathrm{lf}}$ in Eq. (1) is subsequently compared to V_fc, the volume of the segmentation obtained using the initial FC segmentation, in order to estimate the existence of pathology. A significant difference (>T_lf) indicates the presence of potential abnormalities that FC module may have failed to segment. In our implementation, standard statistical type I error rate is considered as the threshold value, >T_lf = 5% of the estimated volume. Any difference within 5% interval of estimated volume is not considered as significant volume difference. However, the smoothness test can still indicate significant changes. Overall, a combined decision based on (i) boundary smoothness and (ii) the estimated lung volumes [see Eq. (2)] triggers the further refinement module of the PLAS system where the subsequent cavity and pathology detection algorithms are run.

$${\displaystyle \underset{\text{smoothness test}}{\underbrace{\{\forall i\in N,C\left(i\right)>K\left(i\right)\}}}\vee \underset{\text{volume difference test}}{\underbrace{\{{\hat{V}}_{\mathrm{lf}}-V_{\mathrm{fc}}\}>T_{\mathrm{lf}}}}\Rightarrow \text{Pathology Presence Test}=\mathrm{True}}$$ (2)

2.C.2. Cavity detection

When the pathology presence test indicates the presence of potential abnormalities in the CT scan, our algorithm automatically initiates a sequence of abnormality detection to segment (i) cavities and (ii) other pathologies.

Cavities occur in multiple lung diseases and are primary markers of TB infection in particular.³⁵ Severe cavities and blebs have boundary with higher contrast separating them from the surrounding lung parenchyma and, therefore, are usually not captured within the initial FC segmentation phase. Because of this, a separate mechanism has been added to the PLAS system to ensure that the cavities are detected and included in the final segmentation. For automatic cavity segmentation, the PLAS system searches the candidate cavity regions automatically by thresholding with strict HU level (< − 990 HU) that denotes gas-filled regions within the lung parenchyma. Since cavity regions are extremely homogeneous and air has a fixed HU level of −1000, only 1% deviation was considered empirically in homogeneity level of those regions when calculating this strict threshold interval (i.e, < − 990 HU). Our search was confined to the regions enclosed by convex-hull, thus eliminating gas-filled regions outside the rib cage (Fig. 4). Next, among all the regions detected as cavities within the convex-hull, we select the voxels that had the minimum HU values as seeds with which we initiated another FC segmentation (m = − 990 HU, σ_ψ = σ_ϕ = − 150 HU). Figure 5 shows a detected cavity region (left), its FC segmentation (middle), and 3D surface rendering (right).

FIG. 5.

Detected cavity (left), the segmentation results using the PLAS method in 2D (middle), and 3D surface rendition (right).

2.C.3. Detection of other abnormal imaging patterns of infectious lung diseases

Following cavity detection, the pathology presence test was repeated again. If the resulting smoothness and volume difference tests still indicated the existence of abnormalities, the machine learning classification at the voxel level was conducted. With voxelwise classification over the boundary areas, it is possible to determine different abnormal imaging patterns. Figure 6 illustrates the voxel classification step in which multiple features from a patch were extracted around the voxel to be classified. The extracted features were subsequently used to classify the voxel as a part of lung or nearby tissues using the random forests predictive learning model obtained in the training part of the machine learning system. Note that the search of classification was restricted within rib cage area.

FIG. 6.

Schematic diagram for machine learning based pathology detection and segmentation.

For random forests voxel classification of lung tissues, we extracted feature sets commonly used in various lung CAD systems: gray-level run length matrix (GLRLM), GLCM, and histogram features. Justification of the use of GLCM, GLRLM, and histogram features was based on a visual appearance of lung pathologies in CT scans.^36,37 Various studies have shown that texture, intensity, and gradient are the key discriminative features for automatically detecting abnormal imaging patterns in human CT scans.^22–27,29 Thus, we designed a patchwise feature set encompassing texture, intensity, and gradient. Briefly, GLCM and GLRLM were calculated using 4-orientations (0°, 45°, 90°, 135°), and for 8,  16,  32,  64,  128 bins.³⁸ For every voxel within the search region, we extracted the features considering a patch (i.e., ROI) of size 5 × 5 centered at the voxels.

Since our aim in this step was to determine the extent of the pathological lung areas missed by earlier segmentation stages of our framework, we considered all pathological regions as a single label (i.e., T_p) in machine learning classification rather than categorization of the abnormality types. Assuming the normal lung parenchyma obtained in the initial delineation step and cavities are represented by ${\mathcal{R}}_{\mathrm{nlp}}$ and ${\mathcal{R}}_{\mathrm{cav}}$, respectively. Then, the pathological regions ${\mathcal{R}}_{\mathrm{rf}}$ can be defined as ${\mathcal{R}}_{\mathrm{rf}}={\mathcal{R}}_{\mathrm{hull}}\setminus ({\mathcal{R}}_{\mathrm{cav}}\hspace{0.17em}\cup \hspace{0.17em}{\mathcal{R}}_{\mathrm{nlp}})$, where ${\mathcal{R}}_{\mathrm{hull}}$ indicates the region enclosed by convex-hull of rib cage. All voxels belonging to ${\mathcal{R}}_{\mathrm{rf}}$ were classified into two classes: pulmonary pathological region (T_p) and neighboring region (T_n). In particular, neighboring structures of the lung were considered as nonlung and labeled as T_n.

For training the RF classifier, two experienced observers annotated various pathology patterns from randomly selected CT scans (21 CT scans from different rabbit and ferret scans). A total of 561 nonoverlapping ROIs were extracted from those annotations among which 237 observations belonged to class T_p while 324 observations belonged to class T_n, and RF was constructed using these observations. For the random forests parameter, we have used 300 trees with depth 10 and 5 predictors sampled at each split.

2.D. Automated airway tree extraction

Once the lung parenchyma has been extracted accurately, the PLAS system initiates the airway extraction module, for which, we propose a method that first enhances airway structures with multiple techniques, then performs delineation based on the enhanced information. From the perspective of appearance, airways can be regarded as local intensity minima surrounded by brighter walls. Such features can be enhanced using 2D gray-scale morphological reconstruction.³⁹ Meanwhile, from the perspective of lumen shape, airways can be treated as dark tubular structures on a bright background. Such structures can be enhanced with vesselness computation using Hessian filtering.⁴⁰ Strengths of these two observations are complementary to each other in the following manner: morphological reconstruction captures local minima without causing false positives, but the enhancement is nonuniform over the image. On the other hand, vesselness computation enhances tubular structures in a more uniform manner, but suffers from false positives along lung boundaries. Therefore, we present a system that combines strengths of these features in a segmentation framework in order to extract the airways accurately. In our approach, the information from both enhancements were incorporated into the affinity function of a multiscale hybrid FC segmentation framework which has ability to combine different affinity functions in a single segmentation problem. The connectivity strength of FC was formulated to fit local airway appearance and shape by combining the complementary information from features including intensity, gray-scale morphological reconstruction, and vesselness enhancement.

For airway segmentation, gray-scale morphological reconstruction and vesselness computation were applied first. To enhance local intensity minima within a 2D image slice, gray-scale morphological reconstruction was performed through a range of 2D morphological structuring elements. The enhancement was performed on a slice plane along all three orientations and the maximum responses are recorded. In the meantime, the vesselness operation was performed to enhance the airway in 3D.

The two enhancement methods above provide complementary information. 2D gray-scale morphological reconstruction does not suffer from false positives, but the degree of enhancement is nonuniform across the airway structure due to the inherent level of enhancement determined by local contrast and 2D operation. For 3D vesselness computation, structure enhancement is homogeneous as the response of matching scale, but false positives are generated over the image, especially along the boundary of lungs. Thus, the two can be combined with intensity and scale information for the formulation of the fuzzy affinity function to facilitate FC computation. In other words, for a given voxel x, we extract three features: intensity G(x), gray-scale morphological reconstructed image D(x), and vesselness measurement V(x). Also, the approximated local scale information, S(x), is available from multiscale vesselness computation. Since intensity is reliable for larger airways while the other two features yield support for smaller ones, instead of using the affinity function using intensity only, we formulated a novel affinity function for airway with FC segmentation,

$${\displaystyle {\mu }_{\psi /\varphi }^A\left(x\right)=\{\begin{array}{ll}{\mu }_{\psi /\varphi }^G\left(x\right),\hfill & \text{if }S\left(x\right)>S_T,\hfill \\ k{\mu }_{\psi /\varphi }^G\left(x\right)+(1-k)\sqrt{{\mu }_{\psi /\varphi }^D\left(x\right){\mu }_{\psi /\varphi }^V\left(x\right)},\hfill & \text{otherwise,}\hfill \end{array}}$$ (3)

where ${\mu }_{\psi /\varphi }^G$, ${\mu }_{\psi /\varphi }^D$, and ${\mu }_{\psi /\varphi }^V$ are affinity functions computed for intensity, morphological reconstructed image, and vesselness features, respectively; S_T ∈ ℝ is the threshold for determining large airway for which the intensity is reliable, and k is the factor to control the ratio of intensity as compared with the other two features in computing the final affinity function ${\mu }_{\psi /\varphi }^{FC}$. It is expected that intensity plays a less important role for finer structures, so k can be formulated as k = S/S_max, where S_max is the local scale controlling parameter.

3. EXPERIMENT AND RESULTS

3.A. Data

The Administrative Panel on Laboratory Animal Care approvals were obtained from each participating institute prior to conducting this research. Internationally recognized standard guidelines were followed as complied with laboratory animal care approvals. To evaluate the performance of the PLAS system, we used two animal models used in infectious lung disease imaging: ferrets infected with pH1N1 and rabbits infected with TB. Two data sets were collected for each animal model. CT images were collected at multiple time points to study the longitudinal progression of the infections.

For the rabbit model, we performed serial CT scans on rabbits infected with the Mycobacterium tuberculosis strain H37Rv. Image acquisition was utilized in a Neurologica CereTom 8 slice CT scanner. A helical scan was collected with the following parameters: 120 keV, 5 mA, and a 0.625 mm slice thickness. Pressure controlled breath-holding was used in one data set to minimize motion artifacts, and it standardized pulmonary pressures for all scans. Nonanesthetized scans were performed in the other data set in a break-proof sealed container with HEPA-filtered gas exchange ports. Animals were only subjected to anesthesia as permitted by their clinical symptoms. For breath-hold scans, rabbit anesthesia was induced in a BSL-3 environment using intramuscular ketamine (20 mg/kg) and xylazine (5–10 mg/kg as required), and rabbits were maintained on 1% isoflurane in 3 l/min of medical oxygen. Animals were kept restrained in a nylon cat bag. These types of cat bags are commonly used by veterinarians and other animal care workers for keeping rabbits stationary for procedures requiring that the animal remain stationary such as blood collection. The animals were incubated and transferred to a custom built chamber in which all joints were sealed and gas exchange occurred through HEPA grade filters. Animals were then transported in this sealed chamber to a BSL-2 environment where CT scanning was performed. Set pressure breath-holds were achieved by closing the expiratory loop of the respiratory circuit to allow the pressure increasing until displacement of a column of water at the relevant depth and then we closed the circuit during acquisition. The duration for which the x-ray source of the CT scanner was active was approximately 15–20 s. For breath-hold scans, the total time that the lungs were inflated was approximately 30 s. The total imaging time, including reconstruction, was about 3–5 min. Rabbits were maintained under gas anesthesia during transport and imaging which totaled 20 min. On average, rabbits would recover to an active and alert conscience state within 2 h of initial anesthesia. Note also that the scans were taken at uniform time points as permitted by the health of the animal. For example, the rabbit (JHU-1) group was imaged every 5 weeks except for the last scan. Euthanasia was required in compliance with the humane treatment of the animals and, therefore, the final scan was at week 38 instead of week 40 as originally scheduled.

For ferrets, imaging was performed on each animal at multiple time points postinfection. Image acquisition was conducted with the Siemens Inveon Trimodal Scanner (Siemens Preclinical Imaging, Knoxville, TN). The Inveon microCT scanner features a variable focus tungsten x-ray source with an achievable resolution of 20 μm and a detector with a maximum field-of-view (FOV) of 8.4 × 5.5 cm. For microCT, the following imaging settings were used: two bed positions, 80 kVp and 500 μA, 500 ms exposure time, and 4 × 4 binning. Ferrets were anesthetized with isoflurane, and general anesthesia was maintained throughout imaging by administering 1%–3% isoflurane via endotracheal tube. Some ferrets were euthanized immediately after imaging to gather virological and histological data to investigate correlations with FDG uptake (see Table I). Two additional ferrets were serially imaged on 1, 2, 3, and 6 days postinfection and euthanized on 7 dpi. Some of the animals were euthanized earlier (e.g., six of the rabbits were sacrificed before week 7) because animal care protocols require that animals be euthanized to limit the amount of suffering experienced by the animals if certain clinical criteria are met. These early terminal endpoints were defined by clinical symptoms such as labored respiration, lack of grooming (symptom of depression), and weight loss. Table I summarizes the data, imaging parameters, and other information regarding the experimental settings.

TABLE I.

Data/imaging parameters.

Animal model	No. of animals	No. of scans	Voxel spacing (mm3)	In-plane resolution	Notes
Rabbits (JHU-1)	10	46	0.3×0.3×0.625	512 × 512	Longitudinal study on TB-infected rabbits. The study was conducted over a period of 38 weeks, six scans were collected at baseline, and 5, 10, 15, 20, and 38 weeks postinfection. One animal was sacrificed at week 5, one animal was sacrificed at week 10, three animals were sacrificed at week 15, and one animal was sacrificed at week 20.
Rabbits (JHU-2)	12	54	0.3×0.3×0.7	512 × 512	Longitudinal study on TB-infected rabbits. The study was conducted over a period of 7 weeks, five scans were collected at baseline, and 3, 4, 5, and 7 weeks postinfection. Six animals were sacrificed at week 5.
Ferrets (KY-1)	8	22	0.21×0.21×0.21	384 × 384	Longitudinal study on pH1N1-infected ferrets. The study was conducted over a period of 5 days, two of the animals have five scans collected at equal intervals, three of them had three scans (first 3 days), and three of them had one scan (first day).
Ferrets (KY-2)	6	11	0.21×0.21×0.21	384 × 384	Longitudinal study on pH1N1-infected ferrets. The study is conducted over a period of 3 days, two scans at 24- and 72-h were collected. One of the animals had one scan (24-h).

3.B. Quantitative and qualitative evaluations

3.B.1. Evaluation of lung segmentation

For the segmentation evaluation of the PLAS, the surrogate truth was provided by two expert observers [“Obs.-I” (observer 1) and “Obs.-II” (observer 2) in Table II]. Our participating experts are radiologists specialized in pulmonary imaging as well as nuclear medicine with more than 10 and 5 yr experience, respectively. Four common quantitative metrics for segmentation accuracy evaluating the agreement between the surrogate truth and the result by the proposed method were used: the Dice Similarity Coefficient (DSC),⁴¹ Hausdorff Distance (HD),⁴¹ sensitivity,⁴¹ and specificity.⁴¹ Table II summarizes the results of quantitative evaluation for all data sets using the proposed system. An average DSC score of 0.9081 was achieved using our method with an average score of 0.9062 for the ferret model, and 0.9087 for the rabbit model. Furthermore, the cavity segmentation was evaluated against manual delineations, and as cavities have relatively high contrast with neighboring tissue, an average DSC score of 0.9807 was achieved. As for pathologies, since the pathologies for infectious disease are distributed across the lung region, manual delineation was not feasible for generating reliable reference truth. Hence, DSC and HD are not applied for pathologies, instead, experts visually validated the pathology region for quantification. However, volume estimation of the pathologies were recorded.

TABLE II.

Overall performance of the proposed lung segmentation approach averaged over different data sets and averaged overall. Mean and standard deviation (std) are provided for each index.

		DSC		HD (mm)		Sensitivity		Specificity
Data set		Obs.-I	Obs.-II	Obs.-I	Obs.-II	Obs.-I	Obs.-II	Obs.-I	Obs.-II
JHU-1	Mean	0.9287	0.9081	25.8487	18.8082	0.9311	0.8781	0.9092	0.9092
	Std	0.0241	0.0531	9.9198	7.5309	0.1213	0.2326	0.1469	0.1335
JHU-2	Mean	0.9028	0.8925	19.8981	14.0238	0.8982	0.8881	0.9097	0.8881
	Std	0.0201	0.1017	12.2223	12.2891	0.1214	0.1887	0.1228	0.1419
KY-1	Mean	0.9162	0.9223	21.8652	19.7654	0.9191	0.8988	0.8584	0.8387
	Std	0.2536	0.1718	9.7827	11.1121	0.0763	0.1278	0.1145	0.1016
KY-2	Mean	0.8881	0.8723	29.2869	28.5781	0.8783	0.9198	0.8718	0.8689
	Std	0.1315	0.1982	18.8875	18.7536	0.0816	0.0835	0.0168	0.1647

For the qualitative assessment, we first showed how PLAS improved the initial region based segmentation. PLAS pathological lung segmentation module was able to include abnormal lung areas that were excluded with FC segmentation as demonstrated in Fig. 7 for rabbit [(A) and (B)] and ferret [(C) and (D)] cases. The comparison was performed on scans with different amount of pathologies. The results obtained using the PLAS system (right) persistently indicated the improved performance over FC method (left).

FIG. 7.

Example axial slices of segmentation results on lung scans of rabbits [(A) and (B)] and ferrets [(C) and (D)] with different pathology. Segmentation results using FC are shown on left, while the results using the PLAS method are presented on the right of each sub-figure.

3.B.2. Statistical comparison and inter- and intraobserver agreements

The PLAS has shown significant improvement in the DSC when compared with FC-only segmentation. Using a paired t-test for the DSCs achieved with the PLAS and the FC, respectively, a statistically significant improvement (p < 0.05) was observed for both ferret and rabbit data sets [see Fig. 8(a) for rabbit data set DSC evaluation as well as inter- and intraobserver agreement rates]. In fact, we do not confine ourselves to the use of the FC for similar research problems; indeed, other successful region-based algorithms can be considered as possible replacement for the FC in the initial segmentation step.

FIG. 8.

(a) Boxplots of DSC for the segmentation evaluation through initial FC segmentation (left) and the proposed PLAS segmentation (right) for rabbit data set. (b) Boxplots of DSC for interobserver (left) and intraobserver agreement rates for rabbit data set (right).

In segmentation evaluation, we followed the standard assessment technique where the two expert observers were asked to repeat the manual delineation after a period of 1 week. The variation between the two delineations was then compared using the DSC values as summarized in Fig. 8(b). We observed that the intra- and interobserver disagreement increases with the increase in the amount of dense pathology and pleura. Hence, the PLAS was found to be very useful in obtaining consistent results among different data sets over time.

3.B.3. Evaluation of airway tree extraction

Qualitatively, an example case of reconstructed airways with the PLAS in comparison with conventional region growing algorithm and its modified version, known as region growing with leakage control, is shown in Fig. 9. As can be noted from the figure, the PLAS has more information about local branches of the airway tree, while it does not have leakage (C). On the other hand, region growing algorithm leaked into nearby tissues because this behavior is often observed particularly when pathological formation exists in the lungs and nearby airway tree structures (A). When leakage control system was used, the thresholding value to separate airway structures from nearby patterns may not be optimal and leading to missing local details. An example for this suboptimal process is illustrated in Fig. 9(B). However, although the leakage problem has been solved (B), many details of the airway branches were lost.

FIG. 9.

Airway lumen segmentation comparison from a rabbit CT scan. Arrow in (A) shows leakage in conventional intensity based region growing algorithm. (B) shows segmented lumen with region growing algorithm when leakage control is used. (C) highlights the proposed method where circles are used to highlight local branches that were not obtained with other algorithms, but with PLAS.

Airway lumen segmentation has natural limitation by the image resolution like other lumen segmentation algorithms. Low resolution can break the continuity of airway structure, especially for smaller airways. Among different algorithms, the proposed method has reasonable segmentation accuracy even when the slice thickness is high by combining enhancements of local structure. In our framework, since we combined two powerful approaches to avoid leakage, our algorithm stopped in higher level of branches instead of leaking on the airway trees. The minimum detectable airway lumen in the PLAS had a width of 2–3 voxels, and, depending on the image spacing, the radius was approximately 0.5–1 mm. This was estimated according to the image resolution of rabbit/ferret. The proposed method was performed on the digitalized image, so the essential limit was reinforced within image space instead of physical space. Therefore, it is expected that the image for smaller animals such as mouse will have smaller voxel size, and thus our method can be translated.

Comparison of algorithms in a broader sense and in quantitative manner is explained in the following. Due to the structural complexity of airway tree, it is extremely difficult and tedious to manually segment airway structure from a CT image. Therefore, we used the data and evaluation criterion provided by EXACT’09 airway segmentation challenge. Some of the initial evaluation results were provided in our preliminary paper.⁴² Detailed descriptions of individual evaluation parameters are available from EXACT’09,⁴³ and multiple evaluation criteria were proposed and utilized to evaluate the performance of candidate algorithms. For a more in-depth evaluation, Table III shows the statistics of the results as compared with the two reference methods (region growing with leakage control and voxel classification) with overall performance. Although EXACT’09 challenge only includes human CT scans, it is the only platform currently providing objective comparison of many airway segmentation algorithms. In this regard, we believe that our algorithm shows strong potential to be used for human data as illustrated and verified by the highly accurate quantitative feedback from the challenge results.

TABLE III.

Results of accuracy analysis of airway segmentation given by EXACT’09 challenge. Statistics with respect to two other participating methods for comparison (UAVisionLab and DIKU) are given for overall performance. EXACT’09 provides an objective evaluation of methods due to availability of ground truth labeling of airway tree.

		Branch count	Branch detected (%)	Tree length (cm)	Tree length detected (%)	Leakage count	Leakage volume (mm3)	False positive rate (FPR) (%)
UAVisionLab	Mean	74.2	32.1	51.9	26.9	4.2	430.4	3.63
	Std	29.5	6.9	19.6	6.9	4.4	672.3	4.92
PLAS	Mean	128.68	51.74	94.81	44.52	8.63	121.37	0.85
	Std	58.68	10.55	43.51	9.14	10.45	284.41	1.54
DIKU	Mean	150.4	59.8	118.4	54.0	1.9	18.2	0.11
	Std	85.2	13.6	75.4	13.4	3.9	48.0	0.22

Based on the “highest tree length” detected under the restriction of small false positive rate <1%, our approach ranked the second best method among all other methods entered into EXACT’09 competition. Visually, we found only “subtle differences” while comparing the top ranked methods, including ours. Additionally, our method exhibited a much higher efficiency (20 min) than the top ranked method (90 min) and does not require a time-consuming training process as the top ranked method does.

In summary, comparing with other state-of-the-art methods, the presented method achieved low false positive rates and the performance was comparable with much higher computational efficiency. As compared with human subjects in EXACT data set, the small animal images used in this paper featured less complex airway structures and higher contrast between airway lumen and lung parenchyma. Therefore, the performance of the proposed method was further enhanced.

3.C. Longitudinal assessment of pulmonary pathologies in infectious lung diseases

Longitudinal imaging of subjects enables tracking of disease progression. Noninvasive longitudinal imaging can accurately assess the lung phenotype, evaluating the global lung measures such as volume, and regional attributes like spatial position and amount of pathology. In convention, the amount of pathology is considered to have linear association with the severity of the disease. The rate of change of pathology volumes over time helps clinicians assess the progression of disease and develop therapy plan if necessary. In this regard, the PLAS system can serve as an effective tool for the pulmonary analysis of longitudinal studies due to its ability to accurately segment normal lung parenchyma, pathological regions, and the airways. To illustrate longitudinal quantification process using PLAS, an example subject [a rabbit infected with TB (JHU-2)] is shown Fig. 10 for three different time points. Segmented lungs, airways, cavities, and other pathologies such as consolidation and GGO are shown in the same plot. Pathology volume progression, their localization, and spatial evolution are illustrated at week 3, 4, and 5 in (A), (B), and (C), respectively. Two slice locations were selected from each week’s rendered surfaces to show in-plane boundary of the airways, cavity, and pathological regions (consolidations and GGO) in the middle column, and their zoomed version in the last column. Additionally, the change of normal lung parenchyma as well as pathological regions can be monitored longitudinally. We further separated the amount of cavity from other pathological regions, due to potential interest in cavity formation,^3,4 and illustrated the change of cavity volumes over time in order to aid clinicians for disease severity assessment. Table IV summarizes the cavity and pathology volume change over time for three rabbits.

FIG. 10.

Longitudinal pulmonary analysis in rabbit data (JHU-2) using the proposed PLAS method. Week 3 (A), week 4 (B), and week 5 (C) were used as time points to show changes in airways, pathological regions, and lung boundaries. Two selected slice locations and their zoomed versions were analyzed in axial domain in the middle and right columns, respectively. In the last column, boundaries that include smaller objects (i.e., airways) indicate mixed formations of GGO and consolidation. Remaining objects in the second, fourth and sixth rows are cavity regions.

TABLE IV.

Cavity volume (ml) over time is shown for three rabbits. Volumetric changes in cavity and other pathologies were found to be significant (>1% − 2% volume change in all cases) given the fact that volumetric agreement between automeasurement and manual measurement of volumes were found to be R² = 0.99 and R² = 0.98 for cavity and other pathologies, respectively.

	Week 3		Week 4		Week 5		Week 7
	Cavity	Pathology	Cavity	Pathology	Cavity	Pathology	Cavity	Pathology
Rabbit C	2.23	3.44	7.12	5.82	4.66	6.89	N/A	N/A
Rabbit H	0.66	4.96	2.56	3.79	2.41	4.23	1.10	4.03
Rabbit I	0.57	3.02	2.41	3.57	2.99	3.72	0.40	3.67

4. DISCUSSIONS

The PLAS system was tested on 133 CT scans from two small animal models (rabbits and ferrets) infected with virus and bacterium (pH1N1 and TB), respectively, and acquired in four different longitudinal studies. By testing rigorously on the versatile data, we made our best effort to cover the most prevalent abnormal imaging patterns in the routine clinical environment. However, additional pre-/postprocessing steps may be needed for certain scans. For instance, cases of extreme noise and motion induced artifacts need to be handled carefully prior to using the PLAS system. Our experiences showed that image data from anesthetized breath-hold animals provided the best image quality with minimal motion artifact. However, there may be other reasons that scientists can use nonanesthetized scans for their quantitative analysis. For instance, anesthesia poses a risk to the animal and, therefore, it would be beneficial to limit the amount of anesthesia administered to the animals. Another added benefit of working with unanesthetized animals is that it is possible to scan 2× more animals in single day. This is mainly due to the time intensive monitoring required as the animals recover from the anesthesia. For other kinds of artifacts, regardless of how robust the FC algorithm, the local descriptors and classifiers, and the rib cage extraction method are, excessive amounts of artifacts may affect the performance of the PLAS system. Nevertheless, appropriate steps can be added prior to using the proposed system.

The proposed method involves the selection of several thresholds. These parameters were chosen adaptively according to the statistics from target structures. The selection is either empirical, such as for normal lung tissue and cavity, or determined from manually defined local ROIs, such as for airway FC image. Details regarding optimal image thresholding method that we followed herein are provided in Ref. 44. To verify the optimization of thresholding process (parameters), expert visual inspections and reference truth from manual delineations were used. Although not observed in the four testing data sets, failures can be corrected by user interaction in cases where it is too difficult to optimize the expected level, due to imaging artifacts such as motion and noise.

One limitation of the PLAS is due to the assumption that rib cage volume is closely associated with lung volume because of the chest anatomy. However, the animal models belonging to phyla lacking those structural characteristics in the chest may not be processed. Despite this limitation, the PLAS still works for most preclinical studies because commonly used animal models for pulmonary diseases have the same structural relationship. In other words, our proposed system can be used in other animal models like mice and rats and for other diseases such as lung cancer and/or COPD due to the similar pathology appearances of those diseases in CT scans. Another limitation of our work is the lack of airway wall thickness assessment and lobe-based analysis of the lungs, which require precise delineation of airway wall and animal specific lobe separation. This will be one of our future tasks as an extension module for the PLAS system.

The proposed method was designed for CT scans in quantifying pulmonary diseases, specifically infections. It has been recently shown that hybrid imaging methods [PET/CT (Ref. 6) and PET/MRI (Ref. 45)] are invaluable for tracking pulmonary infections longitudinally and quantifying metabolic activities as well as structural changes jointly. Incorporation of functional information into our structural pulmonary image analysis platform can be a viable direction in which complementary molecular and anatomical information can be fused for enriched quantification. With the development of new more specific PET-agents for pulmonary infections, we believe that functional quantification will be unique and leading to better understanding of the diseases when combined with the structural markers.

For longitudinal evaluation of both pathology and anatomical structures, a registration approach showing local and global changes could be a useful extension to our module. Since longitudinal quantification of airway surfaces can provide information related to disease progression over time through density, dimension, and shape changes on the wall surface, it could potentially be used as an imaging marker for disease severity evaluation. Currently, our observations in H1N1 and TB showed that higher generation of airways which are closest to the nearby pathologies could potentially be used as severity measurements, but this hypothesis and observation need verification by extensive experiments and evaluation. In case our hypothesis brings development of new imaging marker, related to airway geometry and its associations with disease severity in longitudinal scans, it will be a big breakthrough toward understanding disease severity in infectious disease models using airway geometry.

Toward exploring potential imaging markers in airways for infectious diseases, it should be noted that evaluation of airway tree segmentation in small animal models is a much more complicated task than of organs with simpler shape (liver, bladder, etc.). First, it is quite time-consuming to derive a reference data, mainly due to the complexity of the airway tree structure. Second, it is also challenging to design different evaluation methods for airway analysis, as different levels of airway branches contribute differently for diagnostic purpose. Hence, it would be of interest to develop such a standard data set and evaluation criterion specifically designed for small animal data. This is a significant task of a larger effort in small animal imaging society.

5. CONCLUDING REMARKS

In this study, we presented a novel automated quantitative pulmonary image analysis platform for small animal infectious disease models. While the core segmentation method for lung delineation was based on the fuzzy connectivity algorithm, it was followed by multiple refinement stages in order to handle detection of dense pathologies accordingly. Airways were extracted consecutively using a hybrid multiscale approach based on a novel affinity relationship within the FC framework. The robustness and effectiveness of our approach were tested on 133 CT scans from two widely used small animal infectious disease models.